When GPT-5 thinks like a scientist

The Estonian Subjectivity Dataset has been created to assess document-level subjectivity in the Estonian language, comprising 1,000 documents rated on a scale from 0 (objective) to 100 (subjective) by four annotators. Initial experiments using a large language model (LLM) like GPT-5 for automatic subjectivity analysis showed promising results, although some discrepancies with human annotations were noted.

DeepSeek's recent study highlights the cultural alignment of Large Language Models (LLMs), particularly focusing on how prompt language and cultural prompting affect their outputs. The research utilized Hofstede's VSM13 international surveys to analyze the alignment of models like DeepSeek-V3 and OpenAI's GPT-5 with cultural responses from the United States and China, revealing a significant alignment with the U.S. but not with China.

Google has introduced a new benchmark called 'FACTS' aimed at measuring the factual accuracy of generative AI models, addressing a critical gap in existing benchmarks that focus primarily on task completion rather than the truthfulness of the information generated. This initiative is particularly significant for industries where accuracy is essential, such as legal, finance, and medical sectors.

Tigris has introduced a solution aimed at addressing agent failures in AI systems, which often arise from inconsistent state. The company offers immutable storage, snapshots, and forks to facilitate deterministic and reproducible AI workflows.

A new dataset for Automatic Essay Scoring (AES) and feedback generation in Basque has been introduced, consisting of 3,200 essays annotated by experts. This dataset targets the CEFR C1 proficiency level and includes detailed feedback on various scoring criteria. The study demonstrates that fine-tuning open-source models like Latxa can outperform established systems such as GPT-5 in scoring consistency and feedback quality.

Recent evaluations of advanced reasoning models on mock Chartered Financial Analyst (CFA) exams have shown impressive results, with models like Gemini 3.0 Pro achieving a record score of 97.6% on Level I. This study involved 980 questions across three levels of the CFA exams, and most models successfully passed all levels, indicating a significant improvement in their performance compared to previous assessments of large language models (LLMs).

A new framework named ReasonBreak has been introduced to address privacy concerns associated with multi-modal large reasoning models (MLRMs), which can infer precise geographic locations from personal images using hierarchical reasoning. This framework employs concept-aware perturbations to disrupt the reasoning processes of MLRMs, aiming to enhance geographic privacy protection.

A recent study has demonstrated the potential of large language model (LLM) agents to automate high energy physics data analysis, specifically using the Higgs boson diphoton cross-section measurement as a case study. This hybrid system integrates an LLM-based supervisor-coder agent with the Snakemake workflow manager, allowing for autonomous code generation and execution while ensuring reproducibility and determinism.